Anthropic ships a new Claude generation — the layer under everything we build.

1. 01

   Claude Fable 5 and Mythos 5 land — 1M context, adaptive thinking, new refusal handling

   Anthropic's most capable widely-released models, 1M-token context by default. Our agents run on Claude through the Agent SDK, and the API shape shifted underneath: a dedicated refusal stop-reason, an opt-in fallbacks parameter, and adaptive-thinking-only (manual thinking budgets now error).

2. 02

   Neon makes Postgres 18 the default and 5×'s network transfer

   New projects get Postgres 18 by default; paid-plan public network transfer jumps 100 GB → 500 GB/mo. Neon also previewed three backend services and Lakebase Search. Neon is our database and auth layer, so these defaults and limits are ours.

3. 03

   Vercel Functions now run up to 30 minutes — and Vercel Drop lands

   Node/Python functions can run up to 30 minutes on Pro/Enterprise — more than 2× the old 800-second cap, real headroom for long-running orchestration. Vercel Drop adds drag-and-drop deploys with no Git or CLI. We host on Vercel, so the runtime ceiling is a constraint lifted.

4. 04

   The multiplayer SDK built for exactly our canvas stack

   Not new this week — but newly grounded: Onlook's canvas is React Flow + Liveblocks, and Liveblocks shipped a Multiplayer SDK for React Flow (back in April) that fuses precisely those two. We've never evaluated it. It deserves a real look.

Anthropic released Claude Fable 5 and Mythos 5 on June 9 — its most capable widely-released models, with a 1M-token context window as the default and adaptive thinking always on. There's no manual thinking-budget knob anymore; passing one (or disabled) now errors. The same release reshaped a few API edges: refusals come back with a dedicated stop_reason: "refusal", and a new opt-in fallbacks parameter can re-run a refused request against another model. Alongside, Sonnet 4 and Opus 4 were retired June 15 — those model IDs now error outright.

Why it lands for us: every Onlook project's agent runs on Claude through the Agent SDK, and our own model map already lists claude-fable-5 and claude-mythos-5 as 1M-context options — so these are candidate dispatch models, not distant news. The adaptive-thinking change is the one to verify against our agent loop (anywhere we set a thinking budget needs a look), and the retirements are a reminder to keep an eye on the model-lifecycle cadence even though our current dispatch model isn't among the ones pulled.

Neon made Postgres 18 the default for new projects and raised public network transfer on paid plans from 100 GB to 500 GB per month — a 5× headroom bump on the kind of egress a busy app accrues. It also previewed three coming backend services (Storage, Compute, an AI Gateway) and put Lakebase Search — vector plus BM25 full-text — into private preview.

This is our layer, not adjacent news: Neon is Onlook's database and our auth provider (Neon Auth), so the default Postgres version and the transfer ceiling are decisions that land directly on what we run. The transfer increase is the immediately useful one; Postgres 18 as a default is worth noting for any new project or branch we spin up, and the backend-services preview is the post-acquisition expansion we've been tracking finally taking shape.

Vercel Functions on the Node and Python runtimes can now run up to 30 minutes on Pro and Enterprise — more than double the previous 800-second ceiling. Separately, Vercel Drop lets you deploy a file or folder by dragging it into the browser, with no Git repo, CLI, or local setup, and the Hobby tier's Blob store limit went from 5 to 100.

We host the web app on Vercel and use Vercel Blob for screenshots, uploads, and sandbox backup snapshots, so the runtime ceiling is the one that matters here: long-running orchestration that used to risk the 800-second wall has real headroom now. Vercel Drop is more of a demo-and-prototype convenience than something on our path, and the Blob bump is Hobby-tier — but the function-duration change is a genuine constraint lifted on a surface we run against daily.

A deliberate non-news item, surfaced now because the stack brief made it concrete. Onlook's canvas is built on React Flow, with Liveblocks already powering its realtime presence, cursors, and persistent node state. Back in April, Liveblocks shipped a Multiplayer SDK for React Flow (@liveblocks/react-flow) — a useLiveblocksFlow hook, built-in cursors, and a backend mutateFlow() API — that fuses precisely the two pieces we already operate separately.

We have never evaluated it. That's the whole point of this card: this isn't a thing that happened this week, it's a thing we should have looked at when it landed and didn't. Whether it would simplify our canvas sync or just duplicate what we've hand-wired is an open question — but a provider product built for our exact React-Flow-plus-Liveblocks shape deserves a real spike, not another quarter in the tail.